Unbundling consumer search: query-interface fit

In consumer search, one size does not fit all

 Google search bundled many distinct forms of search into one product:

  • Navigation queries: get to a website
  • Information queries: find a fact
  • Research queries: explore a topic

Google’s interface is a great fit for navigation and information queries, not research queries. The consumer search products I find most compelling exploit this gap in query-interface fit.

Navigation queries

Ten blue links are great for finding your way around the internet. For example, if you search for "YouTube," the first link will always be the YouTube website, and every other link that follows will be relevant. 

Google search results for "Youtube"

Navigation queries are the most common search engine use case and the least difficult to handle. Most search engines would perform equally well. Google's dominance here is partly because the ten blue links interface is well suited for navigation and partly because they spent billions of dollars ensuring every search box on every device you've ever used defaulted to Google. 

Top 5 Google search queries as of April 2024 (YTD)

Information queries

The next most popular type of query is an information query. These are things like searching for the weather or sports scores. Google excels here with specialized interfaces for the most common search terms. 

"Weather" pulls up a mini weather app.

"Premier League" pulls up a comprehensive view with matches, news, standings, stats, and players.

Research queries

The last type of query — research queries — is the least well-served by Google. This is where the opportunity for startups is greatest.

Research queries fall into one of two buckets: depth-first or breadth-first.

Depth-first research queries drive toward a specific answer. For example, in writing this piece, I wanted to know about common categorizations of search queries. ChatGPT was excellent for this — even though some of the categories it came up with were redundant.

I’m looking for a simple, succinct answer in this example, so chat is the right interface.

Another type of research query is breadth-first. I do this often when writing. For example, I may not simply be after a list of search categories but a map of all ideas related to categorizing search queries. It would also be nice if I had the option to go deeper into an idea.

One promising attempt at breadth-first research is Globe Explorer. I think of it as a self-assembling Wikipedia page. It gives you a map of all possible ideas related to your query, organized into a hierarchy that makes it easy to go deep or wide. The results feel exhaustive.

In breadth-first research, you're trying to figure out what questions to ask outside your own knowledge. You may not yet have the right vocabulary to ask the right questions, so having a map of all related topics is exactly the territory you want to explore. A chat interface doesn’t give you a good UX for this type of hierarchical investigation. In the prior example of researching search queries, Globe Explorer outlines types of search queries (what I want) but also related concepts like how a search query is constructed and how exactly it is processed (helpful deep dives). 

One interesting observation is that Globe Explorer uses GPT-3.5 on the backend, so there is not much of a difference in content between Globe Explorer and ChatGPT. The main difference is in the chosen interface: the tree hierarchy view for navigation, the images for details, and the minimal dial-up internet aesthetic (which I found nostalgic). It’s easy to dismiss these as simple interface tweaks, but interface tweaks introduce new mental models — and new mental models are the foundation of entirely new software categories.

Latent space search is another type of breadth-first research I expect to become more common with language models. It’s easiest to understand with an image generation example. In the animation below, each image represents a sampled point in the model’s latent space. Incrementally sampling another point produces a different image.

Panda to Plane

As generative models became popular in 2021, I was convinced latent space exploration was the optimal way to design image creation tools:

 "Given a simple prompt that is less refined than you would like, you can explore the surrounding area of your generated image to find an image that you do, in fact, like. In this implementation, the creative process is visual. Prompts only serve to anchor you to a starting point." 

- Beyond Prompt Engineering

Exploring the surrounding area of a search query is a similar concept that could benefit from a visual creative search tool. I haven’t yet seen a good interface for doing this. The closest existing interface is the graph view implemented by knowledge management tools like Obsidian. This view shows how your ideas are connected. 

A snapshot of an Obsidian local graph. A latent space search interface would let you relate the contents of various nodes in this graph to develop new ideas.

This might be a good starting point for dynamically generating nodes with related ideas.

Brainstorming new product interfaces is also an exercise in brainstorming new business models. Product interfaces anchor non-interface-related things like your monetization strategy. Monetizing clicks in an ad-based revenue model makes sense when your search interface returns a number of links. Query-interface fit comes before product-market fit. What other monetization strategies will emerge from new consumer search interfaces?

It's limiting to think of the search opportunity in AI as disrupting Google. If you're working on a new consumer search product, consider that your opportunity is in unbundling Google and not in disrupting it for queries it is perfectly suited for.